On Mon, 2017-09-18 at 13:21 -0400, John Ferlan wrote: > Rather than checking for whether the devlink.h on the system has > multiple symbols, let's only check for whether the command we want > is defined. > > Turns out the mechanism of providing multiple definitions to check via > AC_CHECK_DECLS in order to determine whether HAVE_DECL_DEVLINK should > be set resulted in a false positive since as long as one of the defs > was true, then the HAVE_DECL_DEVLINK got defined. > > This is further complicated by a change between kernel 4.8 and 4.11 > where the originally defined name DEVLINK_CMD_ESWITCH_MODE_GET was > changed to DEVLINK_CMD_ESWITCH_GET with the comment/caveat that > the old format is obsolete should never be used. Therefore, even > though some kernels will have a couple of the same symbols that > were added at the same time (DEVLINK_ATTR_ESWITCH_MODE and > DEVLINK_ESWITCH_MODE_SWITCHDEV), they won't have the newer name > and thus cause a build failure. > > So even though multiple DEVLINK_CMD_SWITCH* symbols are used to > determine whether the set HAVE_DECL_DEVLINK, this should cover > those kernels before 4.11 with the old definition. > > Signed-off-by: John Ferlan <jferlan(a)redhat.com&gt; > --- > > I'd push this as a build breaker, but I don't want to need to go through > the trouble again if i got this wrong, so hopefully someone who's seeing > this can try out the patch... It's also present on a couple of the CI > infrastructure environments (f23, centos-7): > > https://ci.centos.org/view/libvirt/job/libvirt-master-build/ > > configure.ac | 8 ++++++-- > 1 file changed, 6 insertions(+), 2 deletions(-) > > diff --git a/configure.ac b/configure.ac > index c9509c7f9..73ae7fdd5 100644 > --- a/configure.ac > +++ b/configure.ac > @@ -630,12 +630,16 @@ fi > dnl > dnl check for kernel headers required by devlink > dnl > +dnl kernel 4.8 introduced DEVLINK_CMD_ESWITCH_MODE_GET, but that was > +dnl later changed in kernel 4.11 to be just DEVLINK_CMD_ESWITCH_GET, so > +dnl for "completeness" let's allow HAVE_DECL_DEVLINK to be define if > +dnl either is defined. > if test "$with_linux" = "yes"; then > AC_CHECK_HEADERS([linux/devlink.h]) > - AC_CHECK_DECLS([DEVLINK_GENL_VERSION, DEVLINK_GENL_NAME, DEVLINK_ATTR_MAX, DEVLINK_CMD_ESWITCH_GET, DEVLINK_ATTR_BUS_NAME, DEVLINK_ATTR_DEV_NAME, DEVLINK_ATTR_ESWITCH_MODE, DEVLINK_ESWITCH_MODE_SWITCHDEV], > + AC_CHECK_DECLS([DEVLINK_CMD_ESWITCH_GET, DEVLINK_CMD_ESWITCH_MODE_GET], > [AC_DEFINE([HAVE_DECL_DEVLINK], > [1], > - [whether devlink declarations are available])], > + [whether devlink CMD_ESWITCH_GET is available])], > [], > [[#include <linux/devlink.h>]]) > fi

I was looking into this yesterday, unfortunately I didn't manage to get anything worth posting ready before leaving for the day... Anyway, I don't think your approach will work. First of all, you're removing a number of checks on unrelated symbols that are still used in the code, and if any of those is not present then we shouldn't compile the relevant bits at all.

Second, we're using DEVLINK_CMD_ESWITCH_GET in the code, but as you explain that version is only available in newer kernels. I think the approach need to be more nuanced: - use DEVLINK_CMD_ESWITCH_GET or DEVLINK_CMD_ESWITCH_MODE_GET, depending on which on is available, with the former one being the preferred option;

- if neither DEVLINK_CMD_ESWITCH_GET nor DEVLINK_CMD_ESWITCH_MODE_GET is available, then we should compile the stub; - if any of the other declarations is missing we should also compile the stub. I'll polish up my patch and post it later today.